[Project] Support CAT-Seg from CVPR2023

[SheffieldCao]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Support CAT-Seg open-vocabulary semantic segmentation (CVPR2023).

Modification

Support CAT-Seg open-vocabulary semantic segmentation (CVPR2023).

• Support CAT-Seg model training.
  • CLIP model based backbone (R101 & Swin-B), aggregation layers based neck, and decoder head.
  • Provide customized coco-stuff164k_384x384 training configs.
• Language model supports for open vocabulary (OV) tasks.
  • Support CLIP-based pretrained language model (LM) inference.
  • Add commonly used prompts templates.
• Add README tutorials.
• Add zero-shot testing scripts.

Working on the following tasks.

• Add unit test.

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repos?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
3. If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMDet3D.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[CLAassistant]

All committers have signed the CLA.

[SheffieldCao]

Hi! Is there any update about the review process? @xiexinch

[xiexinch]

Hi! Is there any update about the review process? @xiexinch

I'm working on training these models, but due to the lack of GPU, perhaps we could remove some configs. I'll try to merge it in this week.

[SheffieldCao]

Hi! Is there any update about the review process? @xiexinch

I'm working on training these models, but due to the lack of GPU, perhaps we could remove some configs. I'll try to merge it in this week.

👍Ok, I got it. I have trained these models on my own. The results on COCO dataset in the README is from models trained by me. The zero-short configs of ade and PASCAL are only used for testing. Model is trained only on COCO.
Due to the requirements of pre-commit, I just used model and log links from other models (Mask2Former) as pseudo links. Code Snippet link also need update after the PR is merged.
If there are updates on the mIoU results and configs, please just replace the original ones. Thank you!

[zsc1220]

In the parameter test_class_json，I only filled in two categories, such as cars and trees, but the prediction results show wall and building, which are ade150.json file first two categories, but the categories I need to predict are cars and trees, what should I do? Thank you very much!